static void dumpAndVerifyGraph(Graph& graph, const char* text)
{
GraphDumpMode modeForFinalValidate = DumpGraph;
if (verboseCompilationEnabled()) {
dataLog(text, "\n");
graph.dump();
modeForFinalValidate = DontDumpGraph;
}
if (validationEnabled())
validate(graph, modeForFinalValidate);
}
CompilationResult Plan::finalizeWithoutNotifyingCallback()
{
// We will establish new references from the code block to things. So, we need a barrier.
vm->heap.writeBarrier(codeBlock);
if (!isStillValid()) {
CODEBLOCK_LOG_EVENT(codeBlock, "dfgFinalize", ("invalidated"));
return CompilationInvalidated;
}
bool result;
if (codeBlock->codeType() == FunctionCode)
result = finalizer->finalizeFunction();
else
result = finalizer->finalize();
if (!result) {
CODEBLOCK_LOG_EVENT(codeBlock, "dfgFinalize", ("failed"));
return CompilationFailed;
}
reallyAdd(codeBlock->jitCode()->dfgCommon());
if (validationEnabled()) {
TrackedReferences trackedReferences;
for (WriteBarrier<JSCell>& reference : codeBlock->jitCode()->dfgCommon()->weakReferences)
trackedReferences.add(reference.get());
for (WriteBarrier<Structure>& reference : codeBlock->jitCode()->dfgCommon()->weakStructureReferences)
trackedReferences.add(reference.get());
for (WriteBarrier<Unknown>& constant : codeBlock->constants())
trackedReferences.add(constant.get());
// Check that any other references that we have anywhere in the JITCode are also
// tracked either strongly or weakly.
codeBlock->jitCode()->validateReferences(trackedReferences);
}
CODEBLOCK_LOG_EVENT(codeBlock, "dfgFinalize", ("succeeded"));
return CompilationSuccessful;
}
Plan::CompilationPath Plan::compileInThreadImpl(LongLivedState& longLivedState)
{
if (verboseCompilationEnabled() && osrEntryBytecodeIndex != UINT_MAX) {
dataLog("\n");
dataLog("Compiler must handle OSR entry from bc#", osrEntryBytecodeIndex, " with values: ", mustHandleValues, "\n");
dataLog("\n");
}
Graph dfg(vm, *this, longLivedState);
if (!parse(dfg)) {
finalizer = adoptPtr(new FailedFinalizer(*this));
return FailPath;
}
// By this point the DFG bytecode parser will have potentially mutated various tables
// in the CodeBlock. This is a good time to perform an early shrink, which is more
// powerful than a late one. It's safe to do so because we haven't generated any code
// that references any of the tables directly, yet.
codeBlock->shrinkToFit(CodeBlock::EarlyShrink);
if (validationEnabled())
validate(dfg);
performCPSRethreading(dfg);
performUnification(dfg);
performPredictionInjection(dfg);
if (mode == FTLForOSREntryMode) {
bool result = performOSREntrypointCreation(dfg);
if (!result) {
finalizer = adoptPtr(new FailedFinalizer(*this));
return FailPath;
}
performCPSRethreading(dfg);
}
if (validationEnabled())
validate(dfg);
performBackwardsPropagation(dfg);
performPredictionPropagation(dfg);
performFixup(dfg);
performTypeCheckHoisting(dfg);
unsigned count = 1;
dfg.m_fixpointState = FixpointNotConverged;
for (;; ++count) {
if (logCompilationChanges())
dataLogF("DFG beginning optimization fixpoint iteration #%u.\n", count);
bool changed = false;
if (validationEnabled())
validate(dfg);
performCFA(dfg);
changed |= performConstantFolding(dfg);
changed |= performArgumentsSimplification(dfg);
changed |= performCFGSimplification(dfg);
changed |= performCSE(dfg);
if (!changed)
break;
performCPSRethreading(dfg);
}
if (logCompilationChanges())
dataLogF("DFG optimization fixpoint converged in %u iterations.\n", count);
dfg.m_fixpointState = FixpointConverged;
performStoreElimination(dfg);
// If we're doing validation, then run some analyses, to give them an opportunity
// to self-validate. Now is as good a time as any to do this.
if (validationEnabled()) {
dfg.m_dominators.computeIfNecessary(dfg);
dfg.m_naturalLoops.computeIfNecessary(dfg);
}
switch (mode) {
case DFGMode: {
performTierUpCheckInjection(dfg);
break;
}
case FTLMode:
case FTLForOSREntryMode: {
#if ENABLE(FTL_JIT)
if (FTL::canCompile(dfg) == FTL::CannotCompile) {
finalizer = adoptPtr(new FailedFinalizer(*this));
return FailPath;
}
performCriticalEdgeBreaking(dfg);
performLoopPreHeaderCreation(dfg);
performCPSRethreading(dfg);
performSSAConversion(dfg);
performLivenessAnalysis(dfg);
performCFA(dfg);
performLICM(dfg);
performLivenessAnalysis(dfg);
performCFA(dfg);
performDCE(dfg); // We rely on this to convert dead SetLocals into the appropriate hint, and to kill dead code that won't be recognized as dead by LLVM.
performStackLayout(dfg);
performLivenessAnalysis(dfg);
performFlushLivenessAnalysis(dfg);
performOSRAvailabilityAnalysis(dfg);
dumpAndVerifyGraph(dfg, "Graph just before FTL lowering:");
initializeLLVM();
FTL::State state(dfg);
FTL::lowerDFGToLLVM(state);
if (Options::reportCompileTimes())
beforeFTL = currentTimeMS();
if (Options::llvmAlwaysFailsBeforeCompile()) {
FTL::fail(state);
return FTLPath;
}
FTL::compile(state);
if (Options::llvmAlwaysFailsBeforeLink()) {
FTL::fail(state);
return FTLPath;
}
FTL::link(state);
return FTLPath;
#else
RELEASE_ASSERT_NOT_REACHED();
break;
#endif // ENABLE(FTL_JIT)
}
default:
RELEASE_ASSERT_NOT_REACHED();
break;
}
performCPSRethreading(dfg);
performDCE(dfg);
performStackLayout(dfg);
performVirtualRegisterAllocation(dfg);
dumpAndVerifyGraph(dfg, "Graph after optimization:");
JITCompiler dataFlowJIT(dfg);
if (codeBlock->codeType() == FunctionCode) {
dataFlowJIT.compileFunction();
dataFlowJIT.linkFunction();
} else {
dataFlowJIT.compile();
dataFlowJIT.link();
}
return DFGPath;
}
bool run()
{
ASSERT(m_graph.m_form == SSA);
for (BlockIndex blockIndex = m_graph.numBlocks(); blockIndex--;) {
BasicBlock* block = m_graph.block(blockIndex);
if (!block)
continue;
block->ssa->availabilityAtHead.clear();
block->ssa->availabilityAtTail.clear();
}
BasicBlock* root = m_graph.block(0);
root->ssa->availabilityAtHead.m_locals.fill(Availability::unavailable());
for (unsigned argument = m_graph.m_argumentFormats.size(); argument--;) {
FlushedAt flushedAt = FlushedAt(
m_graph.m_argumentFormats[argument],
virtualRegisterForArgument(argument));
root->ssa->availabilityAtHead.m_locals.argument(argument) = Availability(flushedAt);
}
// This could be made more efficient by processing blocks in reverse postorder.
LocalOSRAvailabilityCalculator calculator(m_graph);
bool changed;
do {
changed = false;
for (BlockIndex blockIndex = 0; blockIndex < m_graph.numBlocks(); ++blockIndex) {
BasicBlock* block = m_graph.block(blockIndex);
if (!block)
continue;
calculator.beginBlock(block);
for (unsigned nodeIndex = 0; nodeIndex < block->size(); ++nodeIndex)
calculator.executeNode(block->at(nodeIndex));
if (calculator.m_availability == block->ssa->availabilityAtTail)
continue;
block->ssa->availabilityAtTail = calculator.m_availability;
changed = true;
for (unsigned successorIndex = block->numSuccessors(); successorIndex--;) {
BasicBlock* successor = block->successor(successorIndex);
successor->ssa->availabilityAtHead.merge(calculator.m_availability);
successor->ssa->availabilityAtHead.pruneByLiveness(
m_graph, successor->at(0)->origin.forExit);
}
}
} while (changed);
if (validationEnabled()) {
for (BlockIndex blockIndex = 0; blockIndex < m_graph.numBlocks(); ++blockIndex) {
BasicBlock* block = m_graph.block(blockIndex);
if (!block)
continue;
calculator.beginBlock(block);
for (unsigned nodeIndex = 0; nodeIndex < block->size(); ++nodeIndex) {
if (block->at(nodeIndex)->origin.exitOK) {
// If we're allowed to exit here, the heap must be in a state
// where exiting wouldn't crash. These particular fields are
// required for correctness because we use them during OSR exit
// to do meaningful things. It would be wrong for any of them
// to be dead.
AvailabilityMap availabilityMap = calculator.m_availability;
availabilityMap.pruneByLiveness(m_graph, block->at(nodeIndex)->origin.forExit);
for (auto heapPair : availabilityMap.m_heap) {
switch (heapPair.key.kind()) {
case ActivationScopePLoc:
case ActivationSymbolTablePLoc:
case FunctionActivationPLoc:
case FunctionExecutablePLoc:
case StructurePLoc:
if (heapPair.value.isDead()) {
dataLogLn("PromotedHeapLocation is dead, but should not be: ", heapPair.key);
availabilityMap.dump(WTF::dataFile());
CRASH();
}
break;
default:
break;
}
}
}
calculator.executeNode(block->at(nodeIndex));
}
}
}
return true;
}
Plan::CompilationPath Plan::compileInThreadImpl(LongLivedState& longLivedState)
{
cleanMustHandleValuesIfNecessary();
if (verboseCompilationEnabled(mode) && osrEntryBytecodeIndex != UINT_MAX) {
dataLog("\n");
dataLog("Compiler must handle OSR entry from bc#", osrEntryBytecodeIndex, " with values: ", mustHandleValues, "\n");
dataLog("\n");
}
Graph dfg(*vm, *this, longLivedState);
if (!parse(dfg)) {
finalizer = std::make_unique<FailedFinalizer>(*this);
return FailPath;
}
codeBlock->setCalleeSaveRegisters(RegisterSet::dfgCalleeSaveRegisters());
// By this point the DFG bytecode parser will have potentially mutated various tables
// in the CodeBlock. This is a good time to perform an early shrink, which is more
// powerful than a late one. It's safe to do so because we haven't generated any code
// that references any of the tables directly, yet.
codeBlock->shrinkToFit(CodeBlock::EarlyShrink);
if (validationEnabled())
validate(dfg);
if (Options::dumpGraphAfterParsing()) {
dataLog("Graph after parsing:\n");
dfg.dump();
}
performLiveCatchVariablePreservationPhase(dfg);
if (Options::useMaximalFlushInsertionPhase())
performMaximalFlushInsertion(dfg);
performCPSRethreading(dfg);
performUnification(dfg);
performPredictionInjection(dfg);
performStaticExecutionCountEstimation(dfg);
if (mode == FTLForOSREntryMode) {
bool result = performOSREntrypointCreation(dfg);
if (!result) {
finalizer = std::make_unique<FailedFinalizer>(*this);
return FailPath;
}
performCPSRethreading(dfg);
}
if (validationEnabled())
validate(dfg);
performBackwardsPropagation(dfg);
performPredictionPropagation(dfg);
performFixup(dfg);
performStructureRegistration(dfg);
performInvalidationPointInjection(dfg);
performTypeCheckHoisting(dfg);
dfg.m_fixpointState = FixpointNotConverged;
// For now we're back to avoiding a fixpoint. Note that we've ping-ponged on this decision
// many times. For maximum throughput, it's best to fixpoint. But the throughput benefit is
// small and not likely to show up in FTL anyway. On the other hand, not fixpointing means
// that the compiler compiles more quickly. We want the third tier to compile quickly, which
// not fixpointing accomplishes; and the fourth tier shouldn't need a fixpoint.
if (validationEnabled())
validate(dfg);
performStrengthReduction(dfg);
performCPSRethreading(dfg);
performCFA(dfg);
performConstantFolding(dfg);
bool changed = false;
changed |= performCFGSimplification(dfg);
changed |= performLocalCSE(dfg);
if (validationEnabled())
validate(dfg);
performCPSRethreading(dfg);
if (!isFTL(mode)) {
// Only run this if we're not FTLing, because currently for a LoadVarargs that is forwardable and
// in a non-varargs inlined call frame, this will generate ForwardVarargs while the FTL
// ArgumentsEliminationPhase will create a sequence of GetStack+PutStacks. The GetStack+PutStack
// sequence then gets sunk, eliminating anything that looks like an escape for subsequent phases,
// while the ForwardVarargs doesn't get simplified until later (or not at all) and looks like an
// escape for all of the arguments. This then disables object allocation sinking.
//
// So, for now, we just disable this phase for the FTL.
//
// If we wanted to enable it, we'd have to do any of the following:
// - Enable ForwardVarargs->GetStack+PutStack strength reduction, and have that run before
// PutStack sinking and object allocation sinking.
// - Make VarargsForwarding emit a GetLocal+SetLocal sequence, that we can later turn into
// GetStack+PutStack.
//
// But, it's not super valuable to enable those optimizations, since the FTL
// ArgumentsEliminationPhase does everything that this phase does, and it doesn't introduce this
// pathology.
changed |= performVarargsForwarding(dfg); // Do this after CFG simplification and CPS rethreading.
}
if (changed) {
performCFA(dfg);
performConstantFolding(dfg);
}
// If we're doing validation, then run some analyses, to give them an opportunity
// to self-validate. Now is as good a time as any to do this.
if (validationEnabled()) {
dfg.ensureDominators();
dfg.ensureNaturalLoops();
dfg.ensurePrePostNumbering();
}
switch (mode) {
case DFGMode: {
dfg.m_fixpointState = FixpointConverged;
performTierUpCheckInjection(dfg);
performFastStoreBarrierInsertion(dfg);
performStoreBarrierClustering(dfg);
performCleanUp(dfg);
performCPSRethreading(dfg);
performDCE(dfg);
performPhantomInsertion(dfg);
performStackLayout(dfg);
performVirtualRegisterAllocation(dfg);
performWatchpointCollection(dfg);
dumpAndVerifyGraph(dfg, "Graph after optimization:");
JITCompiler dataFlowJIT(dfg);
if (codeBlock->codeType() == FunctionCode)
dataFlowJIT.compileFunction();
else
dataFlowJIT.compile();
return DFGPath;
}
case FTLMode:
case FTLForOSREntryMode: {
#if ENABLE(FTL_JIT)
if (FTL::canCompile(dfg) == FTL::CannotCompile) {
finalizer = std::make_unique<FailedFinalizer>(*this);
return FailPath;
}
performCleanUp(dfg); // Reduce the graph size a bit.
performCriticalEdgeBreaking(dfg);
if (Options::createPreHeaders())
performLoopPreHeaderCreation(dfg);
performCPSRethreading(dfg);
performSSAConversion(dfg);
performSSALowering(dfg);
// Ideally, these would be run to fixpoint with the object allocation sinking phase.
performArgumentsElimination(dfg);
if (Options::usePutStackSinking())
performPutStackSinking(dfg);
performConstantHoisting(dfg);
performGlobalCSE(dfg);
performLivenessAnalysis(dfg);
performCFA(dfg);
performConstantFolding(dfg);
performCleanUp(dfg); // Reduce the graph size a lot.
changed = false;
changed |= performStrengthReduction(dfg);
if (Options::useObjectAllocationSinking()) {
changed |= performCriticalEdgeBreaking(dfg);
changed |= performObjectAllocationSinking(dfg);
}
if (changed) {
// State-at-tail and state-at-head will be invalid if we did strength reduction since
// it might increase live ranges.
performLivenessAnalysis(dfg);
performCFA(dfg);
performConstantFolding(dfg);
}
// Currently, this relies on pre-headers still being valid. That precludes running CFG
// simplification before it, unless we re-created the pre-headers. There wouldn't be anything
// wrong with running LICM earlier, if we wanted to put other CFG transforms above this point.
// Alternatively, we could run loop pre-header creation after SSA conversion - but if we did that
// then we'd need to do some simple SSA fix-up.
performLivenessAnalysis(dfg);
performCFA(dfg);
performLICM(dfg);
// FIXME: Currently: IntegerRangeOptimization *must* be run after LICM.
//
// IntegerRangeOptimization makes changes on nodes based on preceding blocks
// and nodes. LICM moves nodes which can invalidates assumptions used
// by IntegerRangeOptimization.
//
// Ideally, the dependencies should be explicit. See https://bugs.webkit.org/show_bug.cgi?id=157534.
performLivenessAnalysis(dfg);
performIntegerRangeOptimization(dfg);
performCleanUp(dfg);
performIntegerCheckCombining(dfg);
performGlobalCSE(dfg);
// At this point we're not allowed to do any further code motion because our reasoning
// about code motion assumes that it's OK to insert GC points in random places.
dfg.m_fixpointState = FixpointConverged;
performLivenessAnalysis(dfg);
performCFA(dfg);
performGlobalStoreBarrierInsertion(dfg);
performStoreBarrierClustering(dfg);
if (Options::useMovHintRemoval())
performMovHintRemoval(dfg);
performCleanUp(dfg);
performDCE(dfg); // We rely on this to kill dead code that won't be recognized as dead by B3.
performStackLayout(dfg);
performLivenessAnalysis(dfg);
performOSRAvailabilityAnalysis(dfg);
performWatchpointCollection(dfg);
if (FTL::canCompile(dfg) == FTL::CannotCompile) {
finalizer = std::make_unique<FailedFinalizer>(*this);
return FailPath;
}
dumpAndVerifyGraph(dfg, "Graph just before FTL lowering:", shouldDumpDisassembly(mode));
// Flash a safepoint in case the GC wants some action.
Safepoint::Result safepointResult;
{
GraphSafepoint safepoint(dfg, safepointResult);
}
if (safepointResult.didGetCancelled())
return CancelPath;
FTL::State state(dfg);
FTL::lowerDFGToB3(state);
if (UNLIKELY(computeCompileTimes()))
m_timeBeforeFTL = monotonicallyIncreasingTimeMS();
if (Options::b3AlwaysFailsBeforeCompile()) {
FTL::fail(state);
return FTLPath;
}
FTL::compile(state, safepointResult);
if (safepointResult.didGetCancelled())
return CancelPath;
if (Options::b3AlwaysFailsBeforeLink()) {
FTL::fail(state);
return FTLPath;
}
if (state.allocationFailed) {
FTL::fail(state);
return FTLPath;
}
FTL::link(state);
if (state.allocationFailed) {
FTL::fail(state);
return FTLPath;
}
return FTLPath;
#else
RELEASE_ASSERT_NOT_REACHED();
return FailPath;
#endif // ENABLE(FTL_JIT)
}
default:
RELEASE_ASSERT_NOT_REACHED();
return FailPath;
}
}
inline bool compile(CompileMode compileMode, ExecState* exec, CodeBlock* codeBlock, JITCode& jitCode, MacroAssemblerCodePtr* jitCodeWithArityCheck, unsigned osrEntryBytecodeIndex)
{
SamplingRegion samplingRegion("DFG Compilation (Driver)");
numCompilations++;
ASSERT(codeBlock);
ASSERT(codeBlock->alternative());
ASSERT(codeBlock->alternative()->getJITType() == JITCode::BaselineJIT);
ASSERT(osrEntryBytecodeIndex != UINT_MAX);
if (!Options::useDFGJIT())
return false;
if (!Options::bytecodeRangeToDFGCompile().isInRange(codeBlock->instructionCount()))
return false;
if (logCompilationChanges())
dataLog("DFG compiling ", *codeBlock, ", number of instructions = ", codeBlock->instructionCount(), "\n");
// Derive our set of must-handle values. The compilation must be at least conservative
// enough to allow for OSR entry with these values.
unsigned numVarsWithValues;
if (osrEntryBytecodeIndex)
numVarsWithValues = codeBlock->m_numVars;
else
numVarsWithValues = 0;
Operands<JSValue> mustHandleValues(codeBlock->numParameters(), numVarsWithValues);
for (size_t i = 0; i < mustHandleValues.size(); ++i) {
int operand = mustHandleValues.operandForIndex(i);
if (operandIsArgument(operand)
&& !operandToArgument(operand)
&& compileMode == CompileFunction
&& codeBlock->specializationKind() == CodeForConstruct) {
// Ugh. If we're in a constructor, the 'this' argument may hold garbage. It will
// also never be used. It doesn't matter what we put into the value for this,
// but it has to be an actual value that can be grokked by subsequent DFG passes,
// so we sanitize it here by turning it into Undefined.
mustHandleValues[i] = jsUndefined();
} else
mustHandleValues[i] = exec->uncheckedR(operand).jsValue();
}
Graph dfg(exec->vm(), codeBlock, osrEntryBytecodeIndex, mustHandleValues);
if (!parse(exec, dfg))
return false;
// By this point the DFG bytecode parser will have potentially mutated various tables
// in the CodeBlock. This is a good time to perform an early shrink, which is more
// powerful than a late one. It's safe to do so because we haven't generated any code
// that references any of the tables directly, yet.
codeBlock->shrinkToFit(CodeBlock::EarlyShrink);
if (validationEnabled())
validate(dfg);
performCPSRethreading(dfg);
performUnification(dfg);
performPredictionInjection(dfg);
if (validationEnabled())
validate(dfg);
performBackwardsPropagation(dfg);
performPredictionPropagation(dfg);
performFixup(dfg);
performTypeCheckHoisting(dfg);
dfg.m_fixpointState = FixpointNotConverged;
performCSE(dfg);
performArgumentsSimplification(dfg);
performCPSRethreading(dfg); // This should usually be a no-op since CSE rarely dethreads, and arguments simplification rarely does anything.
performCFA(dfg);
performConstantFolding(dfg);
performCFGSimplification(dfg);
dfg.m_fixpointState = FixpointConverged;
performStoreElimination(dfg);
performCPSRethreading(dfg);
performDCE(dfg);
performVirtualRegisterAllocation(dfg);
GraphDumpMode modeForFinalValidate = DumpGraph;
if (verboseCompilationEnabled()) {
dataLogF("Graph after optimization:\n");
dfg.dump();
modeForFinalValidate = DontDumpGraph;
}
if (validationEnabled())
validate(dfg, modeForFinalValidate);
JITCompiler dataFlowJIT(dfg);
bool result;
if (compileMode == CompileFunction) {
ASSERT(jitCodeWithArityCheck);
result = dataFlowJIT.compileFunction(jitCode, *jitCodeWithArityCheck);
} else {
ASSERT(compileMode == CompileOther);
ASSERT(!jitCodeWithArityCheck);
result = dataFlowJIT.compile(jitCode);
}
return result;
}
void NaturalLoops::compute(Graph& graph)
{
// Implement the classic dominator-based natural loop finder. The first
// step is to find all control flow edges A -> B where B dominates A.
// Then B is a loop header and A is a backward branching block. We will
// then accumulate, for each loop header, multiple backward branching
// blocks. Then we backwards graph search from the backward branching
// blocks to their loop headers, which gives us all of the blocks in the
// loop body.
static const bool verbose = false;
graph.m_dominators.computeIfNecessary(graph);
if (verbose) {
dataLog("Dominators:\n");
graph.m_dominators.dump(graph, WTF::dataFile());
}
m_loops.resize(0);
for (BlockIndex blockIndex = graph.numBlocks(); blockIndex--;) {
BasicBlock* block = graph.block(blockIndex);
if (!block)
continue;
for (unsigned i = block->numSuccessors(); i--;) {
BasicBlock* successor = block->successor(i);
if (!graph.m_dominators.dominates(successor, block))
continue;
bool found = false;
for (unsigned j = m_loops.size(); j--;) {
if (m_loops[j].header() == successor) {
m_loops[j].addBlock(block);
found = true;
break;
}
}
if (found)
continue;
NaturalLoop loop(successor, m_loops.size());
loop.addBlock(block);
m_loops.append(loop);
}
}
if (verbose)
dataLog("After bootstrap: ", *this, "\n");
FastBitVector seenBlocks;
Vector<BasicBlock*, 4> blockWorklist;
seenBlocks.resize(graph.numBlocks());
for (unsigned i = m_loops.size(); i--;) {
NaturalLoop& loop = m_loops[i];
seenBlocks.clearAll();
ASSERT(blockWorklist.isEmpty());
if (verbose)
dataLog("Dealing with loop ", loop, "\n");
for (unsigned j = loop.size(); j--;) {
seenBlocks.set(loop[j]->index);
blockWorklist.append(loop[j]);
}
while (!blockWorklist.isEmpty()) {
BasicBlock* block = blockWorklist.takeLast();
if (verbose)
dataLog(" Dealing with ", *block, "\n");
if (block == loop.header())
continue;
for (unsigned j = block->predecessors.size(); j--;) {
BasicBlock* predecessor = block->predecessors[j];
if (seenBlocks.get(predecessor->index))
continue;
loop.addBlock(predecessor);
blockWorklist.append(predecessor);
seenBlocks.set(predecessor->index);
}
}
}
// Figure out reverse mapping from blocks to loops.
for (BlockIndex blockIndex = graph.numBlocks(); blockIndex--;) {
BasicBlock* block = graph.block(blockIndex);
if (!block)
continue;
for (unsigned i = BasicBlock::numberOfInnerMostLoopIndices; i--;)
block->innerMostLoopIndices[i] = UINT_MAX;
}
for (unsigned loopIndex = m_loops.size(); loopIndex--;) {
NaturalLoop& loop = m_loops[loopIndex];
for (unsigned blockIndexInLoop = loop.size(); blockIndexInLoop--;) {
BasicBlock* block = loop[blockIndexInLoop];
for (unsigned i = 0; i < BasicBlock::numberOfInnerMostLoopIndices; ++i) {
unsigned thisIndex = block->innerMostLoopIndices[i];
if (thisIndex == UINT_MAX || loop.size() < m_loops[thisIndex].size()) {
insertIntoBoundedVector(
block->innerMostLoopIndices, BasicBlock::numberOfInnerMostLoopIndices,
loopIndex, i);
break;
}
}
}
}
// Now each block knows its inner-most loop and its next-to-inner-most loop. Use
// this to figure out loop parenting.
for (unsigned i = m_loops.size(); i--;) {
NaturalLoop& loop = m_loops[i];
RELEASE_ASSERT(loop.header()->innerMostLoopIndices[0] == i);
loop.m_outerLoopIndex = loop.header()->innerMostLoopIndices[1];
}
if (validationEnabled()) {
// Do some self-verification that we've done some of this correctly.
for (BlockIndex blockIndex = graph.numBlocks(); blockIndex--;) {
BasicBlock* block = graph.block(blockIndex);
if (!block)
continue;
Vector<const NaturalLoop*> simpleLoopsOf;
for (unsigned i = m_loops.size(); i--;) {
if (m_loops[i].contains(block))
simpleLoopsOf.append(&m_loops[i]);
}
Vector<const NaturalLoop*> fancyLoopsOf = loopsOf(block);
std::sort(simpleLoopsOf.begin(), simpleLoopsOf.end());
std::sort(fancyLoopsOf.begin(), fancyLoopsOf.end());
RELEASE_ASSERT(simpleLoopsOf == fancyLoopsOf);
}
}
if (verbose)
dataLog("Results: ", *this, "\n");
}
Plan::CompilationPath Plan::compileInThreadImpl(LongLivedState& longLivedState)
{
Graph dfg(vm, *this, longLivedState);
if (!parse(dfg)) {
finalizer = adoptPtr(new FailedFinalizer(*this));
return FailPath;
}
// By this point the DFG bytecode parser will have potentially mutated various tables
// in the CodeBlock. This is a good time to perform an early shrink, which is more
// powerful than a late one. It's safe to do so because we haven't generated any code
// that references any of the tables directly, yet.
codeBlock->shrinkToFit(CodeBlock::EarlyShrink);
if (validationEnabled())
validate(dfg);
performCPSRethreading(dfg);
performUnification(dfg);
performPredictionInjection(dfg);
if (validationEnabled())
validate(dfg);
performBackwardsPropagation(dfg);
performPredictionPropagation(dfg);
performFixup(dfg);
performTypeCheckHoisting(dfg);
unsigned count = 1;
dfg.m_fixpointState = FixpointNotConverged;
for (;; ++count) {
if (logCompilationChanges())
dataLogF("DFG beginning optimization fixpoint iteration #%u.\n", count);
bool changed = false;
if (validationEnabled())
validate(dfg);
performCFA(dfg);
changed |= performConstantFolding(dfg);
changed |= performArgumentsSimplification(dfg);
changed |= performCFGSimplification(dfg);
changed |= performCSE(dfg);
if (!changed)
break;
performCPSRethreading(dfg);
}
if (logCompilationChanges())
dataLogF("DFG optimization fixpoint converged in %u iterations.\n", count);
dfg.m_fixpointState = FixpointConverged;
performStoreElimination(dfg);
// If we're doing validation, then run some analyses, to give them an opportunity
// to self-validate. Now is as good a time as any to do this.
if (validationEnabled()) {
dfg.m_dominators.computeIfNecessary(dfg);
dfg.m_naturalLoops.computeIfNecessary(dfg);
}
#if ENABLE(FTL_JIT)
if (Options::useExperimentalFTL()
&& compileMode == CompileFunction
&& FTL::canCompile(dfg)) {
performCriticalEdgeBreaking(dfg);
performLoopPreHeaderCreation(dfg);
performCPSRethreading(dfg);
performSSAConversion(dfg);
performLivenessAnalysis(dfg);
performCFA(dfg);
performLICM(dfg);
performLivenessAnalysis(dfg);
performCFA(dfg);
performDCE(dfg); // We rely on this to convert dead SetLocals into the appropriate hint, and to kill dead code that won't be recognized as dead by LLVM.
performLivenessAnalysis(dfg);
performFlushLivenessAnalysis(dfg);
performOSRAvailabilityAnalysis(dfg);
dumpAndVerifyGraph(dfg, "Graph just before FTL lowering:");
// FIXME: Support OSR entry.
// https://bugs.webkit.org/show_bug.cgi?id=113625
FTL::State state(dfg);
FTL::lowerDFGToLLVM(state);
if (Options::reportCompileTimes())
beforeFTL = currentTimeMS();
if (Options::llvmAlwaysFails()) {
FTL::fail(state);
return FTLPath;
}
FTL::compile(state);
FTL::link(state);
return FTLPath;
}
#endif // ENABLE(FTL_JIT)
performCPSRethreading(dfg);
performDCE(dfg);
performVirtualRegisterAllocation(dfg);
dumpAndVerifyGraph(dfg, "Graph after optimization:");
JITCompiler dataFlowJIT(dfg);
if (compileMode == CompileFunction) {
dataFlowJIT.compileFunction();
dataFlowJIT.linkFunction();
} else {
ASSERT(compileMode == CompileOther);
dataFlowJIT.compile();
dataFlowJIT.link();
}
return DFGPath;
}
